PERF: sparse to_csv

asv_bench/benchmarks/io/csv.py

@@ -172,6 +172,26 @@ def time_head_of_multiindex(self):
         self.df_custom_index_then_head.to_csv(self.fname)
 
 
+class ToCSVSparse(BaseIO):
+
+    fname = "__test__.csv"
+
+    def setup(self):
+        from scipy import sparse as sc
+
+        vals = np.random.randint(0, 10, size=(500, 1000))
+        keep = vals > 3
+        vals[keep] = 0
+        sparse_mtx = sc.coo_matrix(vals)
+        self.data = DataFrame.sparse.from_spmatrix(sparse_mtx)
+
+    def time_sparse_to_csv(self):
+        self.data.to_csv(self.fname)
+
+    def time_sparse_to_dense_to_csv(self):
+        self.data.sparse.to_dense().to_csv(self.fname)
+
+
 class StringIORewind:
     def data(self, stringio_object):
         stringio_object.seek(0)

doc/source/whatsnew/v2.0.0.rst

@@ -610,6 +610,7 @@ Performance improvements
 - Performance improvement in :func:`read_stata` with parameter ``index_col`` set to ``None`` (the default). Now the index will be a :class:`RangeIndex` instead of :class:`Int64Index` (:issue:`49745`)
 - Performance improvement in :func:`merge` when not merging on the index - the new index will now be :class:`RangeIndex` instead of :class:`Int64Index` (:issue:`49478`)
 - Performance improvement in :meth:`DataFrame.to_dict` and :meth:`Series.to_dict` when using any non-object dtypes (:issue:`46470`)
+- Performance improvement for saving to CSV with :meth:`DataFrame.to_csv` when data frame is sparse (:issue:`41023`)
 
 .. ---------------------------------------------------------------------------
 .. _whatsnew_200.bug_fixes:

pandas/io/formats/csvs.py

@@ -307,6 +307,11 @@ def _save_chunk(self, start_i: int, end_i: int) -> None:
         slicer = slice(start_i, end_i)
         df = self.obj.iloc[slicer]
 
+        # cast sparse columns to dense to reduce calls
+        # to df._mgr.to_native_types #41023
+        if hasattr(df, "sparse"):
+            df = df.sparse.to_dense()
+
         res = df._mgr.to_native_types(**self._number_format)
         data = [res.iget_values(i) for i in range(len(res.items))]